Project Details
Projekt
Scan2SAM - Data-driven generation of structural systems from 3D point clouds and formalized knowledge for VPINN-based structural forecasting
Subject Area
Geodesy, Photogrammetry, Remote Sensing, Geoinformatics, Cartography
Applied Mechanics, Statics and Dynamics
Applied Mechanics, Statics and Dynamics
Term
since 2022
Project identifier
Deutsche Forschungsgemeinschaft (DFG) - Project number 501834640
The aim of the project is the development and investigation of methods for the automated creation of a structural system from semantically segmented 3D point clouds as well as additional information such as plans, drill cores and expert knowledge. For this purpose, first a structural model and then a structural analysis model by enriching the structural model with boundary conditions will be generated using neuro-symbolic AI. The structural model represents an abstracted geometric-semantic representation of the structure, reflecting its current condition and serving as the foundation for creating the structural analysis model. For this purpose, a spatial graph representation is selected, which is automatically generated from point clouds using deep learning methods. Key aspects of the structural model include the precise capture of the geometric arrangement and the semantic description of the structural components, which are essential for a realistic abstraction of the structure. To enhance the results, novel neuro-symbolic methods are explored, integrating rules and expert knowledge to guide model generation and increase the plausibility of the results. To generate the structural analysis model, the structural model is enriched with additional information such as boundary conditions and internal geometric details. The structural analysis model is characterized by its ability to integrate physical boundary conditions, material properties, and structural relationships in a way that makes it immediately usable for structural analysis. These supplementary details, typically sourced from external data like plans and drill cores, are incorporated into the developed model. The focus is on the automated and explainable transfer of relevant information using clustering and cross-attention (CA) techniques. This results in a complete structural analysis model that serves as the basis for computational methods. Based on the structural analysis model and the monitoring data, the prediction of damaged areas in the structure using Physic-informed Neural Networks (PINNs) and variational PINNs (VPINNs) is investigated. These methods can include measurement data in the forecast in addition to the physical field equations. In the case of existing damage, this also allows stiffness to be included as unknowns in the problem to be solved. On this basis, we want to develop a forecasting tool for structural analysis. Slabs, plates and beams, shear deformable or rigid theories are examined, whereby it is investigated whether PINNs or VPINNs are better suited for structural degradation forecasting from measurement data. These approaches differ in the continuity requirement, which emerges from the underlying theory. An important aspect is the forecast quality in relation to the amount of measurement data needed.
DFG Programme
Priority Programmes
